On 3/9/2017 5:06 PM, Eric Dumazet wrote:
On Thu, 2017-03-09 at 16:35 -0600, Tom Lendacky wrote:
On 3/9/2017 4:17 PM, Jeremy Linton wrote:
Hi,
On 03/09/2017 03:39 PM, Tom Lendacky wrote:
On 3/9/2017 3:26 PM, Jeremy Linton wrote:
Hi,
Hi Jeremy,
I'll have a look at it. Can you send me your kernel config just in
case?
Sure, i will send it to you off list to avoid spamming everyone with a
43k gziped file.
I'm able to recreate the problem with my current config. If I revert
patch 39e6c8208d7b ("net: solve a NAPI race"), then the problem goes
away.
Eric, any thoughts on this? I'm guessing the driver should check that
napi_complete_done returns true before enabling IRQs...
Thanks,
Tom
Yes, it seems this driver has a conditional disable_irq_nosync()
following the napi_schedule_prep()
So it probably needs the same for the enable_irq() side.
I verified that the issue goes away when the return value for
napi_complete_done() is checked. I'll prepare a patch for
both calls in the driver.
This would be broken with busy polling anyway, before my patch.
Yes, I see that busy polling went into 4.10 so I'll submit it back
to 4.10 stable as well.
Thanks,
Tom
diff --git a/drivers/net/ethernet/amd/xgbe/xgbe-drv.c
b/drivers/net/ethernet/amd/xgbe/xgbe-drv.c
index
248f60d171a5a0ce76744a95e2d59039939e6538..c47a3a93f034ab608286d1ae343e9b7778063df8
100644
--- a/drivers/net/ethernet/amd/xgbe/xgbe-drv.c
+++ b/drivers/net/ethernet/amd/xgbe/xgbe-drv.c
@@ -2272,9 +2272,8 @@ static int xgbe_one_poll(struct napi_struct *napi, int
budget)
processed = xgbe_rx_poll(channel, budget);
/* If we processed everything, we are done */
- if (processed < budget) {
- /* Turn off polling */
- napi_complete_done(napi, processed);
+ if (processed < budget &&
+ napi_complete_done(napi, processed)) {
/* Enable Tx and Rx interrupts */
if (pdata->channel_irq_mode)
